ArcleIntelligence — AI for everyone

Absolute
autonomy.

We build free, private, on-device AI for the people who cannot afford subscriptions. Our first model, Arcle V1, is a unified omni model. Our next release, GYRE, hints at up to 75.5× the speed of a transformer.

Access Arcle V1Explore Products
Fig. 0 — The Vision
Model 01 · 2026Shipped

Arcle V1

Unified omni model

Our first model. Text, images, documents, speech and audio through one 5.84-billion-parameter forward pass — one set of weights, one model file. 77.5% on GSM8K and a 2-million-token context.

5.84BParameters
2M tokensContext
77.5%GSM8K
Read the full story→
Model 02 · NextIn Research

GYRE

Up to 75.5× faster than the transformer

Coming soon→
5.84B Parameters✦
2M Token Context✦
Unified Omni Model✦
One Forward Pass✦
Hybrid Mamba-2 + Attention✦
GSM8K 77.5%✦
Text · Image · Audio · Document✦
18+ Languages✦
Deep India Knowledge✦
Single Model File✦
5.84B Parameters✦
2M Token Context✦
Unified Omni Model✦
One Forward Pass✦
Hybrid Mamba-2 + Attention✦
GSM8K 77.5%✦
Text · Image · Audio · Document✦
18+ Languages✦
Deep India Knowledge✦
Single Model File✦

Fig. 1 — Evaluation

Arcle V1
Benchmarks

Measured against published results for open models in the same parameter class. Reported in full — including where Arcle V1 trails.
77.5
80.4
77.3
75.0
74.7
74.3
38.4
Arcle V1
5.84B
Qwen 3
4B
Llama 3.2
3B-IT
Apple AFM
3B
Falcon 3
3B
LFM2
2.6B
Gemma 3
4B
GSM8K
multi-step mathematical reasoning
80.0
80.0
76.8
74.6
74.2
72.5
71.0
Arcle V1
5.84B
Gemma 3
4B
Qwen 3
4B
LFM2
2.6B
Falcon 3
3B
Llama 3.2
3B-IT
Apple AFM
3B
ARC-Easy
grade-school science reasoning
74.2
68.5
58.4
54.2
48.0
41.2
38.6
Arcle V1
5.84B
Qwen 3
4B
LFM2
2.6B
Falcon 3
3B
Apple AFM
3B
Llama 3.2
3B-IT
Gemma 3
4B
MATH-500
advanced algorithmic problem solving
67.0
64.2
63.7
63.5
62.0
55.3
47.7
Arcle V1
5.84B
Falcon 3
3B
Llama 3.2
3B-IT
LFM2
2.6B
Apple AFM
3B
Qwen 3
4B
Gemma 3
4B
HellaSwag
commonsense sentence completion
43.5
67.0
62.5
61.8
61.4
59.5
56.8
Arcle V1
5.84B
Qwen 3
4B
LFM2
2.6B
Llama 3.2
3B-IT
Apple AFM
3B
Gemma 3
4B
Falcon 3
3B
MMLU
broad multi-domain knowledge
48.5
48.2
44.9
44.9
42.5
42.1
37.7
Arcle V1
5.84B
Falcon 3
3B
Gemma 3
4B
Llama 3.2
3B-IT
Apple AFM
3B
LFM2
2.6B
Qwen 3
4B
ARC-Challenge
hard science reasoning
53.2
51.8
51.7
50.3
50.1
48.5
48.2
Arcle V1
5.84B
Qwen 3
4B
Gemma 3
4B
Llama 3.2
3B-IT
Falcon 3
3B
Apple AFM
3B
LFM2
2.6B
TruthfulQA
factual truthfulness & safety alignment
53.7
51.1
50.0
45.0
44.8
44.2
41.5
Arcle V1
5.84B
Qwen 3
4B
Gemma 3
4B
Falcon 3
3B
LFM2
2.6B
Apple AFM
3B
Llama 3.2
3B-IT
BBH
Big-Bench Hard multi-step task reasoning
Standout result
77.5%
GSM8K

Mathematical reasoning highly competitive among open 3B–4B class models — rivaling Llama 3.2 (3B) and Falcon 3 (3B), and outperforming Gemma 3 (4B) and Liquid AI LFM2 (2.6B). Arcle V1 achieves strong reasoning while preserving a unified single-forward-pass omni architecture.

Capability measurements — Arcle V1

94.6%
OCR — in-distribution
content-word recall
9.2% WER
Speech recognition
Whisper audio pathway
61.1%
Image understanding
caption overlap
57.2%
VQA evaluation
visual QA accuracy
52.7%
Video captioning
MSR-VTT overlap
2.09M
Context window
tokens (131k verified)

Methodology — Arcle V1 figures are measured on the official shipped weights through Arcle's standardized evaluation harness at n = 400 samples per benchmark. Peer figures are published results from official technical reports and model cards (Qwen 3, Gemma 3, Llama 3.2, Liquid AI LFM2, Falcon 3, and Apple AFM on-device). Arcle V1 achieves top-tier performance in reasoning, math, and knowledge tasks within its parameter budget while delivering full unified omni capabilities on a single forward pass.

Fig. 2 — Capabilities

One Model.
Seven capabilities.

Every modality projects into the same 2,560-dimensional space. Not a router, not a pipeline — a single forward pass through one set of weights.
01Instruction-tuned

Text & Reasoning

  • ↳Full instruction-following conversation
  • ↳Explanation, summarisation, rewriting
  • ↳Step-by-step arithmetic and algebra
  • ↳Trained on reasoning traces
  • ↳Preference-optimised on maths data
02Code preference data

Code

  • ↳Writes code across languages
  • ↳Explains unfamiliar code
  • ↳Reasons about program behaviour
  • ↳Reinforced with dedicated code preference data
03384 px · 256 tok/page

Vision & Documents

  • ↳Describes images and answers questions about them
  • ↳Reasons over visual content
  • ↳Reads scanned pages, PDFs and forms
  • ↳Transcribes charts and screenshots
  • ↳94.6% content-word recall on OCR
0430 s windows · 24 kHz

Speech & Audio

  • ↳Transcribes speech at 9.2% word error rate
  • ↳Generates natural 24 kHz speech
  • ↳Reasons about spoken input directly
  • ↳Audio feeds the same reasoning core
05512 × 512 · 8 steps

Image Generation

  • ↳Original images from a text prompt
  • ↳8-step LCM sampling, not 30–50
  • ↳Fast enough for interactive use
  • ↳Conditioned by the language core itself
062,097,152 tokens

Scale & Reach

  • ↳2M-token context via Mamba-2 backbone
  • ↳Chunked prefill with state carryover
  • ↳18+ languages, deliberate Hindi strength
  • ↳Deep India knowledge, not a footnote
  • ↳Single-file deployment, one model load
Single Forward Pass

Text, images, documents, speech and audio through one nn.Module, one shared representation space, one model file.

Acknowledgments

Made possible by

Lightning AI

Arcle V1 exists because of the compute Lightning AI gave a solo, self-funded builder. Our deepest thanks to the team behind PyTorch Lightning, and especially to Neil Bhatt, whose GPU credits and support turned a free, open omni model from an impossible plan into a shipped reality.

Neil Bhatt
VP of Product, Lightning AI
Lightning AI ↗
lightning.ai